⚡️ DAILY TECH DIGEST — Sun, Feb 22, 2026

🔬 ML PAPERS (Trending)

🔥 BitDance: Scaling Autoregressive Image Gen with Binary Tokens — AR model using binary diffusion head instead of codebook indices. FID 1.24 on ImageNet, 30x speedup for 1024px gen, 5.4x fewer params than SOTA
📄 https://arxiv.org/abs/2602.14041
💻 https://github.com/shallowdream204/BitDance

🔥 Qwen3-TTS — Multilingual TTS with voice cloning and controllable speech gen. Dual-track LM architecture with specialized speech tokenizers for streaming
📄 https://arxiv.org/abs/2601.15621

🎵 HeartMuLa — Open-source music foundation models. LLM-based song gen with lyrics, style control per section. Claims Suno-level quality with academic-scale resources. Scales to 7B params
📄 https://arxiv.org/abs/2601.10547

🎨 Art2Mus — Direct artwork-to-music generation via visual conditioning. 105K artwork-music pairs dataset. No image-to-text middleman
📄 https://arxiv.org/abs/2602.17599

🖼 RetouchIQ — MLLM agents for instruction-based image retouching with RL + generalist reward model
📄 https://arxiv.org/abs/2602.17558

👗 CORAL — DiT-based virtual try-on with correspondence alignment. Explicit person-garment query-key matching
📄 https://arxiv.org/abs/2602.17636

🎥 GraphThinker — RL finetuning for video reasoning with event graph scene graphs. Reduces hallucinations
📄 https://arxiv.org/abs/2602.17555

🤖 GLM-5: Vibe Coding → Agentic Engineering — Zhipu's new foundation model. DSA for cost reduction, async RL alignment, MIT license
📄 https://arxiv.org/abs/2602.15763